Beruflich Dokumente
Kultur Dokumente
George Phillip
Department
Barker
ofMathemutics
ofMissouri-Kansas
University
Kcmsas City, Missouri 64110
City
ABSTRACT
This survey deals with the aspects of archimedian
partially ordered finitedimensional real vector spaces and order preserving linear maps which do not involve
spectral theory. The first section sketches some of the background
of entrywise
nonnegative
matrices and of systems of inequalities which motivate much of the
current investigations.
The study of inequalities
resulted in the definition
of a
polyhedral cone K and its face lattice F(K). In Section 1I.A the face lattice of a not
necessarily polyhedral cone K in a vector space V is investigated.
In particular the
interplay between the lattice properties of O!(K) and geometric properties of K is
emphasized. Section 1I.B turns to the cones II(K) in the space of linear maps on V.
Recall that II(K) is the cone of all order preserving linear maps. Of particular interest
are the algebraic structure of II(K) as a semiring and the nature of the group Aut( K )
of nonsingular elements A E II( K) for which A E II( K) as well. In a short final
section the cone p,, of n X n positive semidefinite matrices is discussed. A characterization of the set of completely positive linear maps is stated. The proofs will appear in a
forthcoming paper.
I.
HISTORICAL
NOTES
In 1907 Perron [56] initiated the study of matrices A =(aii) all of whose
entries are positive, and Frobenius
[32, 331 carried this study further. If
p=p(A)=max(]Xj:
h is an eigenvalue of A}, then their results showed that p
is an eigenvalue of A which is a simple root of the characteristic
polynomial
and that the eigenvector
belonging to p may be taken to have all entries
positive. In 1912 Frobenius [34] showed that these results hold for a broader
class of matrices with nonnegative
entries. These are the irreducible matrices.
(1981)
263
0024-3795,81/050263
+ 29$02.50
264
PTAP=
AI
A,
then ox+byEK,
We shall also require that K be a closed set. The cone K is fulliff int K # 0,
and K is reproducing iff K-K = V.
REMARK 1.2. In the case of a finite-dimensional V (the situation with
which we shall be principally concerned) a cone K is full iff it is reproducing.
Krein and Rutman considered operators which leave invariant a closed
cone in a Banach space. With certain restrictions on the operators, Krein and
Rutman obtained extensions of the Perron-Frobenius theorems. Birlchoff 1221
gave a proof of the finite-dimensional case of this result. Vandergraft 1721 also
investigated the problem with an eye to the applications to the convergence
of certain iterative processes. These applications have been fruitful, and one
can consult Schrijder [62], Trottenberg and Winter [70], and Vandergraft [73]
for the developments along these lines. The spectral theory of cone preserving
operators is now fairly complete and has been summarized in Barker and
Schneider [14] and Berman and Plemmons [21, Chapter 11. Since the subject
of spectral theory is well summarized, we shall not consider it in detail in the
main body of this paper. However, in order to have a complete extension of
the Perron-Frobenius results it is necessary to have for cone preserving maps
an analog of irreducibility. This extension is based on certain distinguished
subsets of K, namely its faces. The following definition is due to Hans
Schneider.
THEORY OF CONES
265
real
XEK,
y--x~K,and
~EF
imply
XEF.
FaK,
GaK,
Horn(V):
AKCK}.
Then K and II(K) are closed full cones in V and Horn (V)
when K is a closed full cone. Thus f 30 means fE K, and
A E I1( K ). The dimension of a face F, dim F, is the dimension
span F-F of F. If SC K, then Q(S) will denote the least face of
S, viz.,
Q(S)=
respectively
A >O means
of the linear
K containing
FaK}.
266
A polyhedral
theorem).
A cone K is polyhedral
iff it is
The cone K is
and
FvG=@(FUG).
267
THEORY OF CONES
The face lattice of a convex polyhedron is well known (cf. Griinbaum [38] or
McMullen and Shephard [48]). Using the process previously described, we
may translate a theorem about the face lattice of a convex polyhedron into a
theorem about F(K) for a polyhedral cone K and vice versa. We shall say a
great deal more about the lattice F(K) in the next section. For the present we
note one result.
THEOREM 1.9 (Barker [4], Stoer and Witzgall [63]).
5(K)
is a complemented
lattice. Zf K is polyhedral,
Zf K is a cone, then
then 4(K)
is rekztively
complemented,
If K is not polyhedral, then 5(K) is not in general relatively complemented. The reader may refer to Birkhoff [23] for various terms from lattice
theory.
II.
STRUCTURE
THEORY
This section consists of two parts. In the first we consider the structure of
and in the second that of II(K).
However, when it seems
appropriate to do so we shall consider the relationships of K and II(K) in
subsection A. Throughout this section we assume that K is a full closed
pointed cone in V.
K and K,
A.
Let S, T be
268
4(K)
is distributive
iff K is simplicial.
A,x,S-
...+h+lX+l=O.
Zi =
i
hiXi
if
Xi>O,
-Xixi
if
h,<O.
=Zq+l+
. . . +z,.
If zi, zi are distinct extremals, then @(zi)r\Q( xi)=O. Using the distributivity
of %( K ), we have
THEORYOFCONES
269
.--+E)12].
For n=4 this is the Lorentz cone of causal vectors of special relativity. It is
easily seen that every nonzero x E a K is an extremal and consequently ?( K )
is moddar. We can isolate this feature.
DEFINITION2.A.4. Let K be a cone in V. K is strictly convex iff for any
nonzero noncollinear x1, xs ~i3 K we have Q(x,)v@(x,)=K.
It is easily checked that for any strictly convex K, %( K ) is modular.
shall see later, the cone m X m of positive semidefinite matrices is a full
pointed cone in the real space of m Xm hermitian matrices and it
modular face lattice. But for m>2 this cone is not strictly convex.
n=dimV=3
we have the following converse.
THEOREM2.A.5 (Barker [4]).
not distributive,
Problem.
Zf dim K=3
As we
closed
has a
When
but
The genera1 problem is in some sense still open, although we can deal with
some special cases. In another sense the problem is close to finished, as we
shall see later. It is reasonable to consider the polyhedral case first, and in this
situation we have a strong result. Note that for any K the lattice g(K) is of
finite length, so we may take as the definition of semimodularity Corollary 1
of Birkhoff [23, p. 811.
270
If K is polyhedral
and F(K)
is semimodu-
is distributive.
SI
DEFINITION2.A.8.
={fEV:
fs=OVsES}.
fx=OVxEF}.
THEORY OF CONES
271
For FdK where K is polyhedral we have F Ds= F, but for general cones K
this equality does not hold. The next definition singles out this property.
Let K be a cone and FQ K. F is exposed iff F D6= F.
DEFINITION2.A.9.
(i) (F,vF,)~=F~AF$,
(ii) (F, AF,)~Q FfvFzD.
The proof is not difficult and can be found in Barker [7], where examples
are given to show that equality need not hold in (ii). Not surprisingly, equality
in property (ii) relates to whether faces are exposed. In particular we have the
following result.
THEOREM2.A. 11.
(i) (F,AF,)~=F~vF~,
(ii) (G,~G,)s=G~vG~,
F,, F, ET(K),
G,,G,eT(K).
and
As with the polyhedral cones, every face of the Lorentz cone is exposed.
For the Lorentz cone K in R, if we represent the dual space by the same R,
where linear functionals act via the usual inner product, then K and K are
comparable. In fact, K =K. More generally, if V is an inner-product space,
we may identify V with V where f;c= (f, x ) . Whenever we consider R , we
shall assume the usual inner product so that the cones K and K lie in the
same ambient space.
DEFINITION2.A.12.
selfdual iff K = K.
PROPOSITION2.A.13.
Let K be selfdual.
semimodular and every face of K is exposed.
Then g(K)
is modular iff it is
272
This proposition is rather weak, but there are examples of selfdual cones
for which G?(K) is semimodular but not modular and of self-dual polyhedral
cones which are not simplicial. The interesting part of the proof is the
following fact.
LEMMA 2.A.14.
Let K be selfdual.
If FdK,
then FD is a complement
ofF.
In the remainder of subsection A, with a few exceptions, we are concerned only with selfdual K. Consequently, the two duality operators coincide, and we use the notation FD for the dual face. Further, if FdK, then we
let F denote the cone dual of Fin span F. That is,
F =F.
perfect
THEORY OF CONES
whence
273
u) =/3(x,
u#O.
24) +(l-_p)(Y,
u>>P(X> u) >O>
a contradiction,
Hence F = F, and K is perfect.
So far we have not considered the fine structure of cones; specifically,
whether there is any construction
analogous to the direct-sum vector spaces.
Haynsworth [39] considered an external direct sum, but Loewy and Schneider
[46] formulated
the extremely useful notion of decomposability,
which is
analogous to an internal direct sum. Here the cone K need not be self-dual.
DEFINITION 2.A.18.
Let K be a cone in the vector space V. Let K, and
K, be subsets of K. We say K is the direct sum of K, and I(, (and write
K=K,CBK,)
iff
K, = {0},
subsets K, and K,
LEMMA 2.A.19.
Let K==K,@K,.
If XEK,,
~EK,
implies
then (K,@K,)=
K,G3K,. Zf K is selfdual,
then K, = K,.
(x,y)
show
=O,
The proof of the first assertion can be found in Berman [ 191 or Barker and
Foran [12]. The second assertion is discussed in Barker [8]. This lemma is
used in the characterization
of polyhedral perfect cones by Barker and Foran
P21.
THEOREM 2.A.20.
proper maximal face
nonnegative
Zf K is a selfdual
polyhedral cone such that every
is selfdual
in its span, then K is the image of the
transformation.
274
simplicial. Let dim K=n. Let x be an extremal of K, and put F=@(x)~. Then
F is a simplicial cone of dimension n- 1. We claim that K=F@@(
x). Let
K,=F@Q(x).
By Lemma 2.A.19
Thus K,cK
and K,=K,zK=K.
Hence K=K,
n
Theorem (5.3) of Barker [7] is now a corollary of Theorems 2.A.16 and
2.A.20.
COROLLARY2.A.21.
orthomodular
Let K be a selfdual
polyhedral
is
iff K is simplicial.
It should be noted that there are self-dual polyhedral cones which are not
simplicial in any V with dimV>2. Examples and a method of construction
are given in [ 121.
Another corollary of Theorem 2.A.16, this time together with Proposition
2.A.13, is the following result.
COROLLARY2.A.22.
lf K is self-dual
and 4(K)
is modular,
then K is
perfect.
The converse is appealing, since it would classify the self-dual cones with
modular face lattices. The converse is true for n= 1,2,3, but false for n=4.
An example can be found in [9]. To classify modular face lattices, even for
self-dual cones K, we would like to reduce the problem. Since sublattices of
modular lattices are again modular, the next result proves useful.
THEOREM2.A.23 (Barker [B]).
Zf K=K,@K,,
there are faces
then ~(K)=~(K,)G3~(K,).
K,, K, such that $ =OI(Ki),
Conversely, if %(K)=%I@??z,
i= 1,2, and K=K,@K,.
Proof.
The proof of the first statement is a routine verification, so let us
consider the second. If T(K)=91@??S,
then every extremal belongs to either
YTior $,. Let Ki be the join of all extreme rays in i. Since 3(K) is atomic, SO
is Fi, and thus $ = 9(Ki). Similarly K, +K, = K. To finish we must show that
span K, fl span K, = { 0). Obviously, K, n K, = { 0). Suppose z E span K, f?
span K,. For any wi EK,, ui EKE, i= 1,2, for which
.z=wl-w2=u1-u2,
THEORY OF CONES
275
we have
and
it follows that
@(ui)=(a(ua).
(*>
So if
If z#O, then -ZE K,; otherwise -zEK,nspanK,=KznK,={O).
we take ui in the relative interior of K,, there is by Lemma 2.A.17 an E>O
such that ui - EZ= u2 is in the relative boundary of K,. Then
is modular.
Then
9(K)
is subdirectly
irreducible
276
=3.
Zf dim K ~5,
3 and is
strictly convex.
The proof is a tedious argument by cases. We close this section with the
conjecture that the case dim K = 5, h(K) =3 is not possible.
B.
A given closed full cone K completely determines the cones K and Il( K).
In particular, from a knowledge of the extreme rays of K one should be able
to determine the extremals of K and of II( K ). Actually doing this seems to
be quite difficult. Work in this direction was initiated by Loewy and
Schneider [46].
NOTATION 2.B.l.
denoted by Ext K.
Thus the set of all extremals of Il( K) is denoted by Ext n(K).
THEOREM 2.B.2.
are
equivalent:
(1) K is indecomposable;
(2) if A is nonsingular and A(Ext K) CExt K, then A EExt H(K);
(3) if A is nonsingular
(4) ZEExt H(K).
II( K );
Before continuing with the discussion of Ext lX( K) let us give another
interpretation of n( K ). If we represent the action of V on V by (f, x ) and
define an inner product in Hom( V ) by (A, B ) = tr ATB, then we may
identify V@ V with Horn (V), where x@ f( y)= (f, y )x. If K 8 K denotes
the set of all bilinear functionals B such that B( x, f ) 2 0 Vr E K, VIE K , then
K 60 K can be identified with Il( K ). There is the obvious extension to tensor
products of cones K, and K, in vector spaces V, and V,. Here of course if x is
a column vector and f is a row vector, we may identify x@f with the matrix
product xf. Berman and Gaiha [20] noted that n( K ) is the closure of the
convex hull of { f@ x: f E K, x E K }. Tam [67] strengthened this considerably.
THEOREM 2.B.3. II( K ) is equal to the set of all nonnegative
combinations of the form zyr with y E K, -E K.
linear
THEORY OF CONES
277
are equiva-
(1) K is indecomposable,
(2) K is indecomposable,
(3) If( K ) is indecomposable,
(4) A(K) is inclecomposable.
Let xEExt K, fEExt K'. Then x@fEExtII(K)
and f@zrxExtTI(K).
Following the arguments of Haynsworth, Fielder, and Ptak [40], who presented this for polyhedral cones, one can show that every rank-one extremal is
of this form. When are these all?
To phrase this question in terms of tensor products, consider the following
slightly more general situation. Let K, and K, be cones in V, and V,
respectively. As above let K,@KK, denote the set of nonnegative bilinear
functionals on Vi@ VL. That is,
K@K,={B:
B(f,g)>OVfEK;,gEK;}.
Now define
K,@pK,=
{2xi@yi:
xi EK,,
yi EKE},
where all sums are finite. Clearly K,C3 pK, c K,@ K,. Barker [6] showed that
if one of K, and K, is lattice ordered, then equality holds. Further it was
conjectured in that paper that the converse is true. When K, =K; the
statement can be rephrased as a statement about the self-duality of IX(K).
Tam [67] and Barker and Loewy [13] independently settled this question. We
take the statements from the latter paper.
THEOREM 2.B.5.
simplicial.
Let
1 denote
the identity
matrix.
I EA( K)
iff K is
278
For polyhedral cones Fiedler and Pt&k [30, 311 have given a penetrating
analysis of extreme positive operators. Their main result says- roughly speaking-that
the rank of an extreme positive operator may assume any of the
possible values within certain natural boundaries which they spell out, except.
rank two.
If K is polyhedral with Ext K represented by xi,. . ., xv, set P=[r,
...
x,], the nXr matrix with columns xi. We may associate a similar matrix Q
with K. Bums, Fiedler, and Haynsworth [25] discuss the properties of K and
of lI( K) in terms of these matrices.
The situation for the Lorentz cone is also far from trivial. Let C denote the
Lorentz cone in Iw, n&2. Loewy and Schneider [47] established that Ext C
consists of the rank-one extremals together with the set of maps of C onto
itself. Nor is this situation typical. R. C. OBrien [54] has constructed an
indecomposable K C If3r and a nonsingular A E Ext II( K ) which does not take
Ext K into itself.
Clearly a great deal remains to be done in the analysis of lI( K) for
general K.
The collection of onto maps of K are interesting for another reason. They
form a group called the automorphism group of K, Aut( K). Thus A EAut( K)
iff A->O.
A special class of cones arise in the study of automorphic
functions of several complex variables and also in Jordan algebras.
A selfdual cone K CR is homogeneous iff Aut( K)
DEFINITION2.B.7.
acts transitively on int K.
E. B. Vinberg [74] effected a classification of these cones using a correspondence with compact semisimple Jordan algebras. His main result is that
every indecomposable homogeneous cone is unitarily equivalent to a cone of
one of the following classes:
(I) the cone of n X n real positive semidefinite matrices;
(II) the cone of rr X n positive semidefinite hermitian matrices;
(III) the cone of n X n positive semidefinite matrices over the quaternions;
(IV) the cone of 3 X 3 positive semidefinite matrices with elements from
the Cayley numbers;
(V) the Lorentz cone.
(See also Loewy and Schneider [47] for a discussion of the Lorentz cone.) For
a self-dual cone K in a (not necessarily finite-dimensional) real Hilbert space
N there is a weaker notion of homogeneity. If FZI K, it is easy to check that F LJ
is well defined. Let PF denote the orthogonal projection onto F and set
NF=PF-P+.
THEORY OF CONES
279
selfdual
Then a
homogeneous.
compact
group
of
(elementwise)
nonnegative
is finite.
This result does not extend to arbitrary K. For if we take a Lorentz cone
in Iw3, say
and let
0
COSe
-sin0
sin0 ,
cos 6
to a compact
subgroup
of
Proof.
If A is irreducible and has a fixed point in K, then its Perron root
p(A) = 1. Since A E Aut( K ), p( A-) = 1 also, as A has (up to scalar multiples)
only one eigenvector in K. Thus all eigenvalues of A have modulus 1 and
280
linear elementary
divisors (cf. Barker and Schneider
{ A : n = 0, 1,2, . . . } is bounded and hence has compact
[14]).
closure.
This result leads to an extension and alternate proof of the result which
states that a proper Lorentz has an eigenvector in the light cone, that is, the
boundary of a Lorentz cone in R4.
THEOREM 2.B.12.
is even.
in
Proof.
By dividing A by its spectral radius we may assume that A has a
nonzero fixed point in K. If A has no eigenvector in 8 K, it is irreducible (cf.
Vandergraft
[72]). So by Theorem 2.B.11, A belongs to a compact subgroup
of Aut( K ), whence all eigenvalues are of modulus 1 and are simple roots of
the characteristic
equation of A (Vandergraft
[72] or Barker and Schneider
n
[ 141). Since det A >O, then det A = 1. This cannot happen if n is even.
For polyhedral
THEOREM 2.B.13.
Aut( K ) is finite.
Let K be a polyhedral
cone. A compact
group
G C
Let K be an indecomposable
polyhedral
cone.
Then
Aut( K ) is isomorphic to the direct product of the group of positive reals (alias
positive scalar maps)
THEORY OF CONES
281
Proof.
If K is indecomposable,
then by Theorem 2.B.2 ZEE~~ II(K).
Thus the set M of noninvertible
elements of II(K) is closed under addition,
For ZEE~~ II(K) iff some (and hence every) element of Aut( K) is extremal.
Thus the sum of noninvertible
elements is noninvertible.
Therefore M is a
proper (twosided)
ideal of II( K ), and hence the only maximal ideal of any
type. Now suppose K has a nontrivial
decomposition
K = K,@ K,. Define
A,En(K)byA,lspanK,=Z,A,JspanK,=O.DefineAzanalogously.Neither
A, nor A, is invertible, and A,+A,=Z.
Then %(Ai)=Ain(K)
are proper
right ideals, and by a maximality argument each is contained in a maximal
rightideal,say<R(Ai)CL%ifori=1,2.
SinceZ=A,+A2,wehave%,#?Rz,
n
and II( K ) contains at least two distinct maximal ideals.
The cone lI( K) is defined in terms of K, and a vector-space isomorphism
of V clearly induces an isomorphism of lI( K ). To what extent is the converse
trne? Horne [42] considered polyhedral K.
THEOREM2.B.16. Let Vi and V, be finitedimensional
real vector spaces
with full polyhedral cones K, and K, respectively. Suppose T is an isomorphism of the semiring ll( K,) onto H( K,). Then there exists a linear isomorphism
T* : Vi -+ V, mapping
for aZZAETI(K,),
XEK,.
ideal,
II(K
Furthermore,
II(
by each of its nonzero elements.
namely
generated
no nontrivial
minim&
(nonzero)
two-sided
is a principal
two-sided ideal
Also K is simplicial iff II( K) has
two-sided ideals.
Proof.
It is easy to verify that lI( K) is a twosided ideal. Let $l be a
nonzero twosided ideal of II( K ). Choose a nonzero A EP~. Then there are
yEint K, z Eint K such that (-_)rAy=l.
For any ~EK, zEK we have
yzT=(y@jA(yzr)Cf.
Th us $3 II(K
The next statement
is easy to
prove. If II(K) contains no nontrivial twosided ideal, then II( K )=Il( K ).
The last statement now follows from Theorem 2.B.5, since A( K ) = Il( K). n
THEOREM 2.B.18.
II( K,) are isomorphic
Zf the unique
as semirings,
282
THEOREM2.B.19.
n(K).
dimVa2.
scending
sequences
of principal
described
the semiring,
i=l,...,r;
PieAn(K
~,={AEH(K):
Pi en(K)
i=l,...,r.
The proof proceeds along the expected lines except that the proof of closure
under addition employs the fact that the Pi are ertremals in II(K).
ideals in II(A).
THEORY OF CONES
283
Although each -xi, i = 1,. . . , r, is a twosided ideal, not all of these are in
general maximal.
So far we have considered what might be called the arithmetic theory of
ideals. However, ff( K) has another algebraic property, namely the partial
order. The general theory for the faces of II(K) is unsettled at the present.
However, Tam in his thesis and in [69] introduced a simple kind of face.
For F, GET(K)
DEFINITION2.B.22.
set
~I(F,G)=(AEII(K):
AKG}.
Not all faces of fI( K) are of this form, and in particular @( 1 )dff( K) is
not. However, it can be shown that all maximal faces are of this type. Barker
[8] considered the faces fI( F, (0)) and ff( K, F) for a perfect cone (Definition
2.A.15). It is easily seen that fI( F, (0)) is a left ideal and II(K, F) is a right
ideal.
A face of fI( K) which is also a right (left) ideal will
DEFINITION2.B.23.
be called a right (left) facial ideal.
Tam observed that there is a one-to-one correspondence between right
facial ideals of ff( K ) and right ideals in the algebra Hom( V). A similar
remark holds for left and two-sided facial ideals. Thus there are no proper
two-sided facial ideals, although unless K is simplicial, K has proper two-sided
ideals. The ideal structure suggests a classification of the right and left facial
ideals. To verify the details we need some lemmas and notation.
NOTATION2.B.24.
If ScfI(K),
we define ST cII(K)
by
lfF,
G are faces
of K, then
ll( F,G)bI(GD,
In addition
equality
FD).
holds if G is exposed.
Proof
Let AE~(F,G),
~EF,
ZEG~.
Then O=(z,Ay)=(ATz,y),
whence A( G D, c F D. The first statement follows. Assume further that G is
exposed. Taking the transpose of the inclusion, we obtain
lYI( F,G)cIX(G~,F~)?
284
follows.
iff
F,aF,.
Proof. We prove (2), and the proof of (1) is similar. In (2) clearly FlaF2
is true. Let
implies ~(F,,{O})>II(F,,{O}).
So assume the latter containment
y be in the relative interior of FzD, and let x E int K. Recall that x8 y E Il( K)
is defined by (x@y)(z)=(y,
z)x. Then ker(xC3y)nK=Fz
from the.discussion following Definition 2.A.9. Thus
whence
F,aker(x@y)nK=Fz.
COROLLARY 2.B.27.
lhen
THEORY OF CONES
285
form II(F,
(0))
with FaK.
Proof.
To establish the first part note that if @(A) is a left facial ideal of
n(K), then Q(AT)=@(A)r
is a right facial ideal of Il( K). Now modify the
proof of Theorem 2.B.28.
Suppose in addition that F-+ F D is sujective. Then every face of K is
exposed. So from the proof of Theorem 2.B.28 we have
for x lint
Using these results we may also say something about zero divisors in
n(A).
THEOREM2.B.30. A linear operator A EII( K) is a right (or left) zero
divisor iff A belongs to a proper right (or left) facial ideal.
See Tam [69] for the proof.
We close on the refrain of conditions equivalent to K a simplicial cone.
Recall that if K is perfect then every face is exposed.
286
(0)).
Suppose for any F,, F, ET(K) we have that II(F,, {O})+II(F2, (0)) is a left
facial ideal. In particular, if F2 = FF and Fl # {0}, F, #K, then by Theorem
2.B.29 there is a GaK such that
But by Corollary 2.B.27 (I) and Lemma 2.B.26 (l),we have GdF,r\Fro={O},
whence Il(G, {O})=II(K).
Thus there are nonzero Ai EIT(F,, {0}) such that
I =A, + A, and K is decomposable. We check that K = F,@ FID. For any
xEK, x=lx=A~x+A~x.
Further, A<xEFP and A;~EF~, by Lemma 2.B.25
and the fact that FrDD=F. Since it is clear that (span F)n(span FD)= {0},
we have K = F,@ FID.
Now if Fl is any nontrivial face and if x2 EF~ is an extremal, note that
(*>
To verify (*) let ~JEF,~. Then y=axs +x3, where x3 ~@(x,)~,
since K=
@(~,)@@(x,)~.
But then y>x,>O,
so x3 EFF, and thus x3 E@(x,)~/\F~~.
Now start with F, =4)(x1) where x1 is an extremal, and apply (*) n times,
where n = dimV. We obtain K = a( x1) @ * . . $ Ca(x,), and K is simplicial.
The statement for right facial ideals follows from what has been proven
and Lemma 2.B.25.
n
III.
SEMIDEFINITE
MATRICES
Let X, be the real space of nXn hermitian matrices, and let ?inbe the
subset of nXn positive semidefinite matrices. The common inner product on
nXn matrices given by (A, B) = tr(B*A) when restricted to x,, makes x,,
into a real inner product space. It is readily checked that ?i,, is a closed full
pointed cone in X,,. It is also well known (cf. Berman 119, p. 551) that ??nis
self-dual. An outstanding problem in this area is a description of IQ??,,) (cf.
de Pillis [57]). In particular, what are the extremals in II(
As we know
from Sec. II.A, we may identify II(??,,) with Yn@q,,, which is not self-dual,
since Yn is not simplicial (Theorem 2.B.6). First we would like to know the
THEORY OF CONES
287
extreme ray of G?nitself. The classification has been part folklore for a number
of years. The version presented here is from Barker and Carlson [lo], where
there are some indications of other references. Let 8, be the set of orthogonal
projections, that is, A ~(3, iff A ~96, and A2 =A. The usual order on
projections is A< B iff range A crange B. It is easily checked that this order
coincides with the partial order 0, inherits as a subset of Tn. 8, is a lattice if
we define AvB and AAB to be the hermitian projections onto, respectively,
range A + range B and range A n range B. Thus e, is isomorphic with the
lattice of all subspaces Q:n. It is also isomorphic to the face lattice 9(qn) of 9,,.
THEOREM
preserving
lattice isonwrphism.
(thedyadproduct),
(A@,
B)(C)=ACC
The dyad product is the one which corresponds to the tensor products
used in Sec. II. Thus in our present notation we have IT( 9) = qn@n Tn. We
also know from Theorem 2.B.3 that
the inclusions
GEORGE
288
PHILLIP BARKER
THEOREM
3.3. C?,,
is isometrically
matrices.
isomorphic
with
of
n2 X n2 positive semidefinite
Theorem (3.3) indicates why in many respects it is e,, and not II(??,,)
which is the natural notation of a positive operator for maps which preserve
the nXn hermitian matrices.
REFERENCES
6. P. Barker, Monotone norms, Numer. Math. 18:321-326 (1972).
G. P. Barker, On matrices having an invariant cone, Czechoslouak Math. J.
22:49-68 (1972).
G. P. Barker, Matricial norms over cones, Linear Algebra and AppZ. 6:175- 182
(1973).
4 G. P. Barker, The lattice of faces of a finite dimensional cone, Linear Algebra and
Appl. 7:71-82 (1973).
5 G. P. Barker, Stochastic matrices over cones, Linear and Multilinear Algebra
1:279-287 (1974).
6 G. P. Barker, Monotone norms and tensor products, Linear and Multilinear
Algebra 4, No. 3. (1976).
7 G. P. Barker, Faces and duality in convex cones, Linear and M&linear Algebra
6:161-169 (1978).
8 G. P. Barker, Perfect cones, Linear Algebra and AppZ. 22:211-221 (1978).
9 G. P. Barker, Modular face lattices: low dimensional cases, Rocky Mountain J.
Math., to appear.
10 G. P. Barker and D. Carlson, Cones of diagonally dominant matrices, Pacific J.
Math. 57:15-32 (1975).
11 G. P. Barker and D. Carlson, Generalizations of topheavy cones, Linear and
Multilinear Algebra 8:219-230 (1979/80).
12 G. P. Barker and J. Foran, Self-dual cones in Euclidean spaces, Linear Algebra
and Appl. 13: 147- 155 (1976).
13 G. P. Barker and R. Loewy, The structure of cones of matrices, Linear Algebra
and Appl. 12:87-94 (1975).
14 G. P. Barker and H. Schneider, Algebraic Perron-Frobenius theory, Linear
Algebra and AppZ. 11:219-233 (1975).
15 G. P. Barker and R. E. L. Turner, Some observations on the spectra of cone
preserving maps, Linear Algebra and Appl. 6:149- 15.3 (1973).
16 F. L. Bauer, Lecture Notes on Positivity and Norms; Part I: Ordering and
Positiuity, Technische Universitgt Miinchen, (1974).
17 J. Bellissard and B. Iochum, Homogeneous self dual cones versus Jordan algebras,
Ann. ht. Fourier 28:104-144 (1978).
18 J. Bellissard, B. Iochum, and R. Lima, Homogeneous and facially homogeneous
self-dual cones, Linear Algebra and Appl. 19:1- 16 (1978).
THEORY
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
OF CONES
289
290
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
GEORGE
PHILLIP
BARKER
THEORY
OF CONES
291
65
66
B.-S. Tam, Some results on cross-positive matrices, Linear Algebra and Appl.
15:173- 176 (1976).
B.-S. Tam, Some results on polyhedral cones and simplicial cones, Linear and
Mu&linear Algebra 4:281-284 (1977).
B.-S. Tam, On indecomposable
subcones, Linear Algebra and A&. 22:33-40
(1978).
B.-S. Tam, On the semiring of cone preserving maps, Linear Algebra and Appl.
35:79- 108 (1981).
U. Trottenberg
and G. Winter, On faces and W-irreducible operators in finite
dimensional linear spaces, Linear Algebra and Appl. 33:87- 110 (1980).
H. Uzawa, A theorem on convex polyhedral cones, in Studies in Linear and
Nonlinear Progrumming (K. J. Arrow, L. Hurwicz, and H. Uzawa, Eds.), Stanford
U. P., Stanford, Calif., (1958).
J. S. Vandergraft,
Spectral properties of matrices which have invariant cones,
SIAM J. Appl. Math. 16:1208-1222 (1968).
J. S. Vandergraft, A note on irreducibility for linear operators on partially ordered
finite dimensional vector spaces, Linear Algebra and AppE. 13:139- 146 (1976).
E. B. Vinberg, The theory of convex homogeneous cones, Trans. Moscow Math.
Sot. 12:340-403 (1963).
H. Wielandt, Unzerlebare nicht-negative
Matrizen, Math. Zeitsch. 52:642-648
(1956).
67
68
69
70
71
72
73
74
75
cones, J. Austral.